<p align="center"> <img src="https://raw.githubusercontent.com/cvs-health/langfair/main/assets/images/langfair-logo.png" /> </p>

LangFair: Use-Case Level LLM Bias and Fairness Assessments

LangFair is a comprehensive Python library designed for conducting bias and fairness assessments of large language model (LLM) use cases. This repository includes various supporting resources, including

• Documentation site with complete API reference
• Comprehensive framework for choosing bias and fairness metrics
• Demo notebooks providing illustrative examples
• LangFair tutorial on Medium
• Software paper on how LangFair compares to other toolkits
• Research paper on our evaluation approach

🚀 Why Choose LangFair?

Static benchmark assessments, which are typically assumed to be sufficiently representative, often fall short in capturing the risks associated with all possible use cases of LLMs. These models are increasingly used in various applications, including recommendation systems, classification, text generation, and summarization. However, evaluating these models without considering use-case-specific prompts can lead to misleading assessments of their performance, especially regarding bias and fairness risks.

LangFair addresses this gap by adopting a Bring Your Own Prompts (BYOP) approach, allowing users to tailor bias and fairness evaluations to their specific use cases. This ensures that the metrics computed reflect the true performance of the LLMs in real-world scenarios, where prompt-specific risks are critical. Additionally, LangFair's focus is on output-based metrics that are practical for governance audits and real-world testing, without needing access to internal model states.

<p align="center"> <img src="https://raw.githubusercontent.com/cvs-health/langfair/release-branch/v0.4.0/assets/images/langfair_graphic.png" /> </p>

Note: This diagram illustrates the workflow for assessing bias and fairness in text generation and summarization use cases.

⚡ Quickstart Guide

(Optional) Create a virtual environment for using LangFair

We recommend creating a new virtual environment using venv before installing LangFair. To do so, please follow instructions here.

Installing LangFair

The latest version can be installed from PyPI:

pip install langfair

Usage Examples

Below are code samples illustrating how to use LangFair to assess bias and fairness risks in text generation and summarization use cases. The below examples assume the user has already defined a list of prompts from their use case, prompts.

Generate LLM responses

To generate responses, we can use LangFair's ResponseGenerator class. First, we must create a langchain LLM object. Below we use ChatVertexAI, but any of LangChain’s LLM classes may be used instead. Note that InMemoryRateLimiter is to used to avoid rate limit errors.

from langchain_google_vertexai import ChatVertexAI
from langchain_core.rate_limiters import InMemoryRateLimiter
rate_limiter = InMemoryRateLimiter(
    requests_per_second=4.5, check_every_n_seconds=0.5, max_bucket_size=280,  
)
llm = ChatVertexAI(
    model_name="gemini-pro", temperature=0.3, rate_limiter=rate_limiter
)

We can use ResponseGenerator.generate_responses to generate 25 responses for each prompt, as is convention for toxicity evaluation.

from langfair.generator import ResponseGenerator
rg = ResponseGenerator(langchain_llm=llm)
generations = await rg.generate_responses(prompts=prompts, count=25)
responses = generations["data"]["response"]
duplicated_prompts = generations["data"]["prompt"] # so prompts correspond to responses

Compute toxicity metrics

Toxicity metrics can be computed with ToxicityMetrics. Note that use of torch.device is optional and should be used if GPU is available to speed up toxicity computation.

# import torch # uncomment if GPU is available
# device = torch.device("cuda") # uncomment if GPU is available
from langfair.metrics.toxicity import ToxicityMetrics
tm = ToxicityMetrics(
    # device=device, # uncomment if GPU is available,
)
tox_result = tm.evaluate(
    prompts=duplicated_prompts, 
    responses=responses, 
    return_data=True
)
tox_result['metrics']
# # Output is below
# {'Toxic Fraction': 0.0004,
# 'Expected Maximum Toxicity': 0.013845130120171235,
# 'Toxicity Probability': 0.01}

Compute stereotype metrics

Stereotype metrics can be computed with StereotypeMetrics.

from langfair.metrics.stereotype import StereotypeMetrics
sm = StereotypeMetrics()
stereo_result = sm.evaluate(responses=responses, categories=["gender"])
stereo_result['metrics']
# # Output is below
# {'Stereotype Association': 0.3172750176745329,
# 'Cooccurrence Bias': 0.44766333654278373,
# 'Stereotype Fraction - gender': 0.08}

Generate counterfactual responses and compute metrics

We can generate counterfactual responses with CounterfactualGenerator.

from langfair.generator.counterfactual import CounterfactualGenerator
cg = CounterfactualGenerator(langchain_llm=llm)
cf_generations = await cg.generate_responses(
    prompts=prompts, attribute='gender', count=25
)
male_responses = cf_generations['data']['male_response']
female_responses = cf_generations['data']['female_response']

Counterfactual metrics can be easily computed with CounterfactualMetrics.

from langfair.metrics.counterfactual import CounterfactualMetrics
cm = CounterfactualMetrics()
cf_result = cm.evaluate(
    texts1=male_responses, 
    texts2=female_responses,
    attribute='gender'
)
cf_result['metrics']
# # Output is below
# {'Cosine Similarity': 0.8318708,
# 'RougeL Similarity': 0.5195852482361165,
# 'Bleu Similarity': 0.3278433712872481,
# 'Sentiment Bias': 0.0009947145187601957}

Alternative approach: Semi-automated evaluation with AutoEval

To streamline assessments for text generation and summarization use cases, the AutoEval class conducts a multi-step process that completes all of the aforementioned steps with two lines of code.

from langfair.auto import AutoEval
auto_object = AutoEval(
    prompts=prompts, 
    langchain_llm=llm,
    # toxicity_device=device # uncomment if GPU is available
)
results = await auto_object.evaluate()
results['metrics']
# # Output is below
# {'Toxicity': {'Toxic Fraction': 0.0004,
#   'Expected Maximum Toxicity': 0.013845130120171235,
#   'Toxicity Probability': 0.01},
#  'Stereotype': {'Stereotype Association': 0.3172750176745329,
#   'Cooccurrence Bias': 0.44766333654278373,
#   'Stereotype Fraction - gender': 0.08,
#   'Expected Maximum Stereotype - gender': 0.60355167388916,
#   'Stereotype Probability - gender': 0.27036},
#  'Counterfactual': {'male-female': {'Cosine Similarity': 0.8318708,
#    'RougeL Similarity': 0.5195852482361165,
#    'Bleu Similarity': 0.3278433712872481,
#    'Sentiment Bias': 0.0009947145187601957}}}

Bias and Fairness Red-Teaming

To assess worst-case toxicity and counterfactual generations for a given use case, LangFair also offers off-the-shelf red-teaming evaluations. The following code can be used:

from langfair.generator import AdversarialGenerator
ag = AdversarialGenerator(langchain_llm=llm)

# Generate responses to adversarial prompts (toxicity)
toxicity_generations = await ag.toxicity() 

# Generate responses to adversarial prompts (counterfactual fairness)
counterfactual_generations = await ag.counterfactual(group_categories=["Gender", "Race/ethnicity"],)

📚 Example Notebooks

Explore the following demo notebooks to see how to use LangFair for various bias and fairness evaluation metrics:

• Toxicity Evaluation: A notebook demonstrating toxicity metrics.
• Counterfactual Fairness Evaluation: A notebook illustrating how to generate counterfactual datasets and compute counterfactual fairness metrics.
• Stereotype Evaluation: A notebook demonstrating stereotype metrics.
• AutoEval for Text Generation / Summarization (Toxicity, Stereotypes, Counterfactual): A notebook illustrating how to use LangFair's AutoEval class for a comprehensive fairness assessment of text g